Make the checkpoint's dspark_target_layer_ids authoritative for DSpark
speculative decoding. When target_layer_ids is unset it is copied from
the checkpoint verbatim (order preserved, since main_proj columns are
order-dependent). An explicit override that does not exactly match the
checkpoint list is now rejected during config validation instead of
failing later with a projection shape mismatch (different count) or
silently degrading acceptance (same count, different/reordered layers).

Add unit coverage for defaulting, matching override, mismatched count,
same-count-different-layers, and order-mismatch cases.

Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com>

Rebase DSpark onto the unified one-model speculative-decoding base introduced
by PR NVIDIA#15775 and route production and acceptance through the shared entry
points instead of hand-rolling them in DSparkWorker:

- Produce draft tokens via SpecWorkerBase.sample_draft_tokens on the per-position
  corrected block logits [num_gens, K, vocab] (surfaced from the draft model via
  the existing return_logits path), letting the base do greedy/rejection sampling,
  TP gather, d2t, and the draft_probs scatter.
- Accept via SpecWorkerBase.sample_and_accept_draft_tokens (strict or rejection),
  dropping the local _sample_and_accept_draft_tokens_base call.
- Fill context requests draft-prob rows with a one-hot via
  write_context_onehot_draft_probs.
- Register DSpark in the rejection gate (is_supported / is_new_rejection_method)
  and thread vocab_size / draft_vocab_size / use_rejection_sampling into
  DSparkSpecMetadata so the slot-indexed rejection buffers are allocated.

Greedy behavior is preserved (the base argmaxes the same corrected logits the
head produced). Adds a mixed-batch (context + gen) worker orchestration test.

Validated: 60/60 hw_agnostic unit tests; 8xB300 e2e (BENCH_OK); and the full
DL1-7 thinking on/off DEP8+MegaMoE-DeepGEMM AL sweep matches the previously
collected baseline within N=12 nondeterministic variance.

Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com>

The optional real-fp8-MLA draft-attention path (env TLLM_DSPARK_REAL_MLA /
DSparkDraftModel.use_real_mla) was opt-in, never used by any production or
benchmark config (all use the default bf16 captured-context path), and is
currently broken: it runs torch.einsum/baddbmm on Float8_e4m3fn tensors, which
torch does not implement ("baddbmm_cuda not implemented for Float8_e4m3fn").
It is not on the roadmap, so remove it rather than ship dead/broken code.

Removed:
- _dspark_mla_attention() and the TLLM_DSPARK_REAL_MLA env / use_real_mla attr
  and all use_real_mla branches in modeling_dspark.py (attention dispatch,
  weight-load setup, and the forward_batched NotImplementedError guard).
- The eager per-request DSparkWorker._draft_gen_block fallback (which existed
  only for real-fp8-MLA) and the _use_batched toggle; the worker is now always
  the batched, CUDA-graph-safe path.
- The obsolete legacy-real-mla worker test.

The default bf16 batched path is unchanged (byte-identical, just un-nested).
Validated: 59/59 DSpark hw_agnostic unit tests pass (worker/draft/heads/
attention/cuda_graph). Net -256 lines.

Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com>

… width

Address PR review (r3601639329). DSparkWorker.forward passed
gen_logits.shape[-1] to write_context_onehot_draft_probs, but DSpark surfaces
vocab-sharded draft logits (unlike DFlash, which gathers in the worker), so
under tensor parallelism that is the pre-gather shard width.
sample_draft_tokens all-gathers the logits to full vocab and scatters
draft_probs at that full width (publishing draft_probs_last_dim), so writing the
context requests' one-hot at the shard width leaves the trailing columns stale
and corrupts rejection acceptance on context-to-gen transitions.

Use spec_metadata.draft_probs_last_dim (the width the gen scatter just
published) for the context one-hot. Update the mixed-batch worker test to assert
the one-hot uses the scatter width, not the sharded gen_logits width.

Only affects the rejection-sampling path under plain TP (no-op for greedy and
gated off under attention-DP); validated by the DSpark hw_agnostic unit suite
(59 passed).

Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com>

…onfig

Confidence-based dynamic drafting is intentionally not enabled in this PR (the
draft path always calls with confidence_threshold=0.0). Keep the confidence head
as scaffolding for a follow-up, but address the related review feedback and
remove the user-facing knobs that are not wired up yet:

- Gate DSparkConfidenceHead(with_markov=...) on markov_rank > 0 (was hardcoded
  True). build_markov_head() returns None for markov_rank <= 0, and
  dspark_propose then passes no prev_embeddings, which would trip the
  prev_embeddings assert in DSparkConfidenceHead.forward (review r3502819934).
- Remove the unwired DSparkDecodingConfig fields enable_confidence_head and
  confidence_threshold (read nowhere at runtime; review r3547358128). They will
  be re-added when the confidence head is actually wired into the speculative
  scheduler/verifier. Regenerate the LLM args golden manifest accordingly.
- Document in dspark_propose that num_proposed is not yet consumed and the
  confidence branch is inert scaffolding (review r3502819941).

The confidence head module and its internal plumbing remain in place. No runtime
behavior change (the confidence path was already never exercised).
Validated: DSpark hw_agnostic unit suite (24 passed) and test_llm_args DSpark
tests (9 passed).

Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com>